File size: 8,535 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 |
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
$assert ROW_TILE >= 1
$assert ACCUMULATORS >= 1
#include <assert.h>
#include <wasm_simd128.h>
#include <xnnpack/dwconv.h>
#include <xnnpack/math.h>
$ARCH_SUFFIX = "_x86" if X86 else "_arm"
void xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd${ARCH_SUFFIX}_loadsplat_${ROW_TILE}x4${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}(
size_t input_height,
size_t input_width,
const float* input,
const float* weights,
const float* zero,
float* output,
uint32_t padding_top,
const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(input_height != 0);
assert(input_width != 0);
assert(input_width % sizeof(float) == 0);
assert(padding_top == 1);
const v128_t vmask = wasm_v128_load(params->wasmsimd_stride1.mask);
const v128_t vmax = wasm_v128_load64_splat(params->wasmsimd_stride1.max);
const v128_t vmin = wasm_v128_load64_splat(params->wasmsimd_stride1.min);
const v128_t vw0123 = wasm_v128_load(weights);
const v128_t vw4567 = wasm_v128_load(weights + 4);
const v128_t vw89 = wasm_v128_load64_splat(weights + 8);
const v128_t vbias = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0);
const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1);
const v128_t vk01 = wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2, 2);
const v128_t vk02 = wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3, 3);
const v128_t vk10 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0);
const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1);
const v128_t vk12 = wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2, 2);
const v128_t vk20 = wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3, 3);
const v128_t vk21 = wasm_v32x4_shuffle(vw89, vw89, 0, 0, 0, 0);
const v128_t vk22 = wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1);
const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
const float* i0 = zero;
const float* i1 = input;
$for M in range(2, 2 + ROW_TILE):
const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);
float* o0 = output;
$for M in range(1, ROW_TILE):
float* o${M} = (float*) ((uintptr_t) o${M-1} + input_width);
size_t output_height = input_height;
do {
$for M in range(2, 2 + ROW_TILE):
if XNN_UNPREDICTABLE(output_height < ${M}) {
i${M} = zero;
$if M <= ROW_TILE:
o${M-1} = o${M-2};
}
$for M in range(2 + ROW_TILE):
v128_t vi${M}x0123 = wasm_f32x4_const_splat(0.0f);
$for M in range(2 + ROW_TILE):
v128_t vi${M}x4567 = wasm_v128_load(i${M});
i${M} += 4;
size_t w = input_width;
for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
$for M in range(2 + ROW_TILE):
const v128_t vi${M}x89AB = wasm_v128_load(i${M});
i${M} += 4;
$for K in range(3):
$for M in range(ROW_TILE):
$if K == 0:
v128_t vo${M}p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi${M+K}x4567, vk${K}1));
$elif K < ACCUMULATORS:
v128_t vo${M}p${K} = wasm_f32x4_mul(vi${M+K}x4567, vk${K}1);
$else:
vo${M}p${K % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${K % ACCUMULATORS}, wasm_f32x4_mul(vi${M+K}x4567, vk${K}1));
$for M in range(2 + ROW_TILE):
const v128_t vi${M}x3456 = wasm_v32x4_shuffle(vi${M}x0123, vi${M}x4567, 3, 4, 5, 6);
$for K in range(3):
$for M in range(ROW_TILE):
$if K+3 < ACCUMULATORS:
v128_t vo${M}p${K+3} = wasm_f32x4_mul(vi${M+K}x3456, vk${K}0);
$else:
vo${M}p${(K+3) % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${(K+3) % ACCUMULATORS}, wasm_f32x4_mul(vi${M+K}x3456, vk${K}0));
$for M in range(2 + ROW_TILE):
vi${M}x0123 = vi${M}x4567;
$for M in range(2 + ROW_TILE):
const v128_t vi${M}x5678 = wasm_v32x4_shuffle(vi${M}x4567, vi${M}x89AB, 1, 2, 3, 4);
$for K in range(3):
$for M in range(ROW_TILE):
vo${M}p${(K+6) % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${(K+6) % ACCUMULATORS}, wasm_f32x4_mul(vi${M+K}x5678, vk${K}2));
$for M in range(2 + ROW_TILE):
vi${M}x4567 = vi${M}x89AB;
$if ACCUMULATORS > 1:
$ACC_SLICE = 1
$while ACC_SLICE < ACCUMULATORS:
$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
$if A + ACC_SLICE < ACCUMULATORS:
$for M in range(ROW_TILE):
vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
$ACC_SLICE *= 2
$if X86:
$for M in range(ROW_TILE):
v128_t vo${M} = wasm_f32x4_pmax(vmin, vo${M}p0);
$for M in range(ROW_TILE):
vo${M} = wasm_f32x4_pmin(vmax, vo${M});
$else:
$for M in range(ROW_TILE):
v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
$for M in range(ROW_TILE):
vo${M} = wasm_f32x4_min(vo${M}, vmax);
$for M in reversed(range(ROW_TILE)):
wasm_v128_store(o${M}, vo${M});
o${M} += 4;
}
// Always process the last block of 1..4 pixels.
assert(w >= 1 * sizeof(float));
assert(w <= 4 * sizeof(float));
{
$for M in range(2 + ROW_TILE):
vi${M}x4567 = wasm_v128_and(vmask, vi${M}x4567);
$for K in range(3):
$for M in range(ROW_TILE):
$if K == 0:
v128_t vo${M}p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi${M+K}x4567, vk${K}1));
$elif K < ACCUMULATORS:
v128_t vo${M}p${K} = wasm_f32x4_mul(vi${M+K}x4567, vk${K}1);
$else:
vo${M}p${K % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${K % ACCUMULATORS}, wasm_f32x4_mul(vi${M+K}x4567, vk${K}1));
$for M in range(2 + ROW_TILE):
const v128_t vi${M}x3456 = wasm_v32x4_shuffle(vi${M}x0123, vi${M}x4567, 3, 4, 5, 6);
$for K in range(3):
$for M in range(ROW_TILE):
$if K+3 < ACCUMULATORS:
v128_t vo${M}p${K+3} = wasm_f32x4_mul(vi${M+K}x3456, vk${K}0);
$else:
vo${M}p${(K+3) % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${(K+3) % ACCUMULATORS}, wasm_f32x4_mul(vi${M+K}x3456, vk${K}0));
const v128_t vzero = wasm_f32x4_const_splat(0.0f);
$for M in range(2 + ROW_TILE):
const v128_t vi${M}x5678 = wasm_v32x4_shuffle(vi${M}x4567, vzero, 1, 2, 3, 4);
$for K in range(3):
$for M in range(ROW_TILE):
vo${M}p${(K+6) % ACCUMULATORS} = wasm_f32x4_add(vo${M}p${(K+6) % ACCUMULATORS}, wasm_f32x4_mul(vi${M+K}x5678, vk${K}2));
$if ACCUMULATORS > 1:
$ACC_SLICE = 1
$while ACC_SLICE < ACCUMULATORS:
$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
$if A + ACC_SLICE < ACCUMULATORS:
$for M in range(ROW_TILE):
vo${M}p${A} = wasm_f32x4_add(vo${M}p${A}, vo${M}p${A + ACC_SLICE});
$ACC_SLICE *= 2
$if X86:
$for M in range(ROW_TILE):
v128_t vo${M} = wasm_f32x4_pmax(vmin, vo${M}p0);
$for M in range(ROW_TILE):
vo${M} = wasm_f32x4_pmin(vmax, vo${M});
$else:
$for M in range(ROW_TILE):
v128_t vo${M} = wasm_f32x4_max(vo${M}p0, vmin);
$for M in range(ROW_TILE):
vo${M} = wasm_f32x4_min(vo${M}, vmax);
if XNN_LIKELY(w == 4 * sizeof(float)) {
$for M in reversed(range(ROW_TILE)):
wasm_v128_store(o${M}, vo${M});
o${M} += 4;
} else {
if (w & (2 * sizeof(float))) {
$for M in reversed(range(ROW_TILE)):
wasm_v128_store64_lane(o${M}, vo${M}, 0);
o${M} += 2;
$for M in range(ROW_TILE):
vo${M} = wasm_v64x2_shuffle(vo${M}, vo${M}, 1, 1);
}
if (w & (1 * sizeof(float))) {
$for M in reversed(range(ROW_TILE)):
wasm_v128_store32_lane(o${M}, vo${M}, 0);
o${M} += 1;
}
}
}
i0 = (const float*) ((uintptr_t) i${ROW_TILE} - input_decrement);
i1 = (const float*) ((uintptr_t) i${ROW_TILE+1} - input_decrement);
$for M in range(2, 2 + ROW_TILE):
i${M} = (const float*) ((uintptr_t) i${M-1} + input_width);
$if ROW_TILE > 1:
o0 = o${ROW_TILE - 1};
$for M in range(1, ROW_TILE):
o${M} = (float*) ((uintptr_t) o${M-1} + input_width);
$if ROW_TILE > 1:
output_height = doz(output_height, ${ROW_TILE});
} while (${"--" if ROW_TILE == 1 else ""}output_height != 0);
}
|